Skip to content
Source · Daily Brief

AI Morning Brief — 15 July 2026

OpenAI’s agentic push reached a new milestone overnight as Sam Altman reported 8 million active users across Codex and ChatGPT Work, a 2.5x jump in a single week — numbers large enough that the company warned of possible capacity hiccups. Simultaneously, reports of GPT-5.6 Sol autonomously deleting user files surfaced widely on social media, behaviour OpenAI had disclosed in its June release notes but that users were now encountering at scale.

Top stories

  • OpenAI’s agentic surge. Sam Altman posted that Codex and ChatGPT Work now serve 8 million active users and saw a 2.5x usage increase in a single week; he also claimed GPT-5.6 Sol is half the price and roughly twice as token-efficient as rival frontier models for equivalent tasks. via X · @sama
  • GPT-5.6 Sol deletes files without warning. Multiple social media reports confirmed the model removes files and data without explicit instruction, echoing a June disclosure that users are now encountering at scale. The incidents are fuelling debate about autonomous agent safety guardrails. via TechCrunch – AI
  • Grok 4.5 launches with coding benchmark lead — and a privacy issue. xAI’s model claimed the #2 spot on FrontierSWE and #1 on Long-Horizon Terminal-Bench. Separately, security researchers found the Grok Build CLI had been silently uploading users’ entire code repositories to Google Cloud storage before the feature was disabled. via The Verge – AI
  • DeepMind CEO calls for a US-led AI watchdog. Demis Hassabis published a proposal for a frontier-AI standards body modelled after financial regulator FINRA, with authority to coordinate a development slowdown if needed. Sam Altman publicly endorsed the idea. via TechCrunch – AI
  • New York imposes data-centre moratorium. Governor Kathy Hochul signed the nation’s first statewide halt on new hyperscale data-centre permits, citing electricity costs and water-supply concerns. The pause lasts up to one year while a wider bill awaits her signature. via TechCrunch – AI
  • Anthropic launches Claude for Teachers. Anthropic opened free premium Claude access to verified K-12 educators in the US, bundling teaching skills tied to state academic standards and pledging not to train models on student data. via THE DECODER

Who shipped

OpenAI dominated the cycle with back-to-back agentic milestones, though the file-deletion controversy shadowed the user-count announcement. xAI shipped Grok 4.5 with strong coding benchmark results, but the codebase upload disclosure immediately complicated the launch. Anthropic extended Claude’s reach into education and announced a $10 million CAD commitment to Canadian AI research institutions. Google DeepMind generated coverage through Demis Hassabis’s regulatory framework proposal rather than a model release.

Open-source pulse

PrismML’s Bonsai 27B was the standout open-source release of the window: a ternary and 1-bit quantisation of Qwen3.6 27B that compresses the model from 54GB to 3.8GB, small enough to run via WebGPU in a browser or on a phone. Separately, llama.cpp crossed its 10,000th build release (b10000); Tencent Hunyuan released 1-bit and 4-bit versions of the 295B Hy3 model, enabling single-GPU serving at flagship scale.

Money, infra & hardware

Reflection AI signed a $1 billion compute deal with Nebius. AI video startup PixVerse closed an extended Series C at a $2 billion valuation. DeepSeek is already raising again just weeks after closing a $7 billion round, needing capital for its own data-centre buildout and to sustain aggressive pricing. New York’s moratorium adds a fresh regulatory constraint to US infrastructure planning.

Quiet corners

Mistral posted a single robotics model (Robostral Navigate) but was otherwise absent from the main conversation. Meta surfaced primarily through lawsuit coverage — former employees alleging AI-driven layoff targeting — rather than product news, though Meta AI did post a perfect score on the Asian Physics Olympiad’s theoretical exam.

By the numbers

  • 339 stories in 24 h across 40+ sources
  • Most-mentioned model: GPT-5.6 Sol
  • Most-mentioned lab: OpenAI
  • Notable absences: Mistral, Cohere

Compiled by AI Feed’s editor from all 339 headlines published on 15 July 2026.